| | import math |
| | import os |
| |
|
| | from scipy import integrate |
| | import torch |
| | from torch import nn |
| | import torchsde |
| | from torchdiffeq import odeint |
| | from tqdm.auto import trange, tqdm |
| | from matplotlib import pyplot as plt |
| | import numpy as np |
| |
|
| | from . import utils |
| |
|
| |
|
| | ''' |
| | Code adapted for state-action based sampling with/without goal-conditioning: |
| | |
| | https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py |
| | ''' |
| |
|
| | def append_zero(action): |
| | return torch.cat([action, action.new_zeros([1])]) |
| |
|
| |
|
| | def get_sigmas_karras(n, sigma_min, sigma_max, rho=7., device='cpu'): |
| | """Constructs the noise schedule of Karras et al. (2022).""" |
| | ramp = torch.linspace(0, 1, n) |
| | min_inv_rho = sigma_min ** (1 / rho) |
| | max_inv_rho = sigma_max ** (1 / rho) |
| | sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho |
| | return append_zero(sigmas).to(device) |
| |
|
| |
|
| | def get_sigmas_exponential(n, sigma_min, sigma_max, device='cpu'): |
| | """Constructs an exponential noise schedule.""" |
| | sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), n, device=device).exp() |
| | return append_zero(sigmas) |
| |
|
| |
|
| | def get_sigmas_linear(n, sigma_min, sigma_max, device='cpu'): |
| | """Constructs an linear noise schedule.""" |
| | sigmas = torch.linspace(sigma_max, sigma_min, n, device=device) |
| | return append_zero(sigmas) |
| |
|
| |
|
| | def cosine_beta_schedule(n, s=0.008, device='cpu'): |
| | """ |
| | cosine schedule |
| | as proposed in https://openreview.net/forum?id=-NEXDKk8gZ |
| | """ |
| | steps = n + 1 |
| | x = np.linspace(0, steps, steps) |
| | alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2 |
| | alphas_cumprod = alphas_cumprod / alphas_cumprod[0] |
| | betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1]) |
| | betas_clipped = np.clip(betas, a_min=0, a_max=0.999) |
| | return append_zero(torch.tensor(np.flip(betas_clipped).copy(), device=device, dtype=torch.float32)) |
| |
|
| |
|
| | def get_sigmas_ve(n, sigma_min=0.02, sigma_max=100, device='cpu'): |
| | """Constructs a continuous VP noise schedule.""" |
| | |
| | steps = n + 1 |
| | t = torch.linspace(0, steps, n, device=device) |
| | t = (sigma_max ** 2) * ((sigma_min ** 2 / sigma_max ** 2) ** (t / (n - 1))) |
| | sigmas = torch.sqrt(t) |
| | return append_zero(sigmas) |
| |
|
| |
|
| | def get_iddpm_sigmas(n, sigma_min=0.02, sigma_max=100, M=1000, j_0=0, C_1=0.001, C_2=0.008, device='cpu'): |
| | """Constructs a continuous VP noise schedule.""" |
| | |
| | step_indices = torch.arange(n, dtype=torch.float64, device=device) |
| | u = torch.zeros(M + 1, dtype=torch.float64, device=device) |
| | alpha_bar = lambda j: (0.5 * np.pi * j / M / (C_2 + 1)).sin() ** 2 |
| | for j in torch.arange(M, j_0, -1, device=device): |
| | u[j - 1] = ((u[j] ** 2 + 1) / (alpha_bar(j - 1) / alpha_bar(j)).clip(min=C_1) - 1).sqrt() |
| | u_filtered = u[torch.logical_and(u >= sigma_min, u <= sigma_max)] |
| | sigmas = u_filtered[((len(u_filtered) - 1) / (n - 1) * step_indices).round().to(torch.int64)] |
| | return append_zero(sigmas).to(torch.float32) |
| |
|
| |
|
| | def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'): |
| | """Constructs a continuous VP noise schedule.""" |
| | t = torch.linspace(1, eps_s, n, device=device) |
| | sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1) |
| | return append_zero(sigmas) |
| |
|
| |
|
| | def to_d(action, sigma, denoised): |
| | """Converts a denoiser output to a Karras ODE derivative.""" |
| | return (action- denoised) / utils.append_dims(sigma, action.ndim) |
| |
|
| |
|
| |
|
| | def default_noise_sampler(x): |
| | return lambda sigma, sigma_next: torch.randn_like(x) |
| |
|
| |
|
| |
|
| | def get_ancestral_step(sigma_from, sigma_to, eta=1.): |
| | """Calculates the noise level (sigma_down) to step down to and the amount |
| | of noise to add (sigma_up) when doing an ancestral sampling step.""" |
| | if not eta: |
| | return sigma_to, 0. |
| | sigma_up = min(sigma_to, eta * (sigma_to ** 2 * (sigma_from ** 2 - sigma_to ** 2) / sigma_from ** 2) ** 0.5) |
| | sigma_down = (sigma_to ** 2 - sigma_up ** 2) ** 0.5 |
| | return sigma_down, sigma_up |
| |
|
| |
|
| | class BatchedBrownianTree: |
| | """A wrapper around torchsde.BrownianTree that enables batches of entropy.""" |
| |
|
| | def __init__(self, x, t0, t1, seed=None, **kwargs): |
| | t0, t1, self.sign = self.sort(t0, t1) |
| | w0 = kwargs.get('w0', torch.zeros_like(x)) |
| | if seed is None: |
| | seed = torch.randint(0, 2 ** 63 - 1, []).item() |
| | self.batched = True |
| | try: |
| | assert len(seed) == x.shape[0] |
| | w0 = w0[0] |
| | except TypeError: |
| | seed = [seed] |
| | self.batched = False |
| | self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed] |
| |
|
| | @staticmethod |
| | def sort(a, b): |
| | return (a, b, 1) if a < b else (b, a, -1) |
| |
|
| | def __call__(self, t0, t1): |
| | t0, t1, sign = self.sort(t0, t1) |
| | w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign) |
| | return w if self.batched else w[0] |
| |
|
| |
|
| | class BrownianTreeNoiseSampler: |
| | """A noise sampler backed by a torchsde.BrownianTree. |
| | Args: |
| | x (Tensor): The tensor whose shape, device and dtype to use to generate |
| | random samples. |
| | sigma_min (float): The low end of the valid interval. |
| | sigma_max (float): The high end of the valid interval. |
| | seed (int or List[int]): The random seed. If a list of seeds is |
| | supplied instead of a single integer, then the noise sampler will |
| | use one BrownianTree per batch item, each with its own seed. |
| | transform (callable): A function that maps sigma to the sampler's |
| | internal timestep. |
| | """ |
| |
|
| | def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x): |
| | self.transform = transform |
| | t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max)) |
| | self.tree = BatchedBrownianTree(x, t0, t1, seed) |
| |
|
| | def __call__(self, sigma, sigma_next): |
| | t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next)) |
| | return self.tree(t0, t1) / (t1 - t0).abs().sqrt() |
| |
|
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_euler( |
| | model, |
| | state: torch.Tensor, |
| | action: torch.Tensor, |
| | goal: torch.Tensor, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | s_churn=0., |
| | s_tmin=0., |
| | s_tmax=float('inf'), |
| | s_noise=1. |
| | ): |
| | """ |
| | Implements a variant of Algorithm 2 (Euler steps) from Karras et al. (2022). |
| | Stochastic sampler, which combines a first order ODE solver with explicit Langevin-like "churn" |
| | of adding and removing noise. |
| | Every update consists of these substeps: |
| | 1. Addition of noise given the factor eps |
| | 2. Solving the ODE dx/dt at timestep t using the score model |
| | 3. Take Euler step from t -> t+1 to get x_{i+1} |
| | |
| | In contrast to the Heun variant, this variant does not compute a 2nd order correction step |
| | For S_churn=0 the solver is an ODE solver |
| | """ |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0. |
| | eps = torch.randn_like(action) * s_noise |
| | sigma_hat = sigmas[i] * (gamma + 1) |
| | |
| | if gamma > 0: |
| | action = action + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5 |
| | denoised = model(state, action, goal, sigma_hat * s_in, **extra_args) |
| | d = to_d(action, sigma_hat, denoised) |
| | if callback is not None: |
| | callback({'x': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised}) |
| | dt = sigmas[i + 1] - sigma_hat |
| | |
| | action = action + d * dt |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_euler_ancestral( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | eta=1. |
| | ): |
| | """ |
| | Ancestral sampling with Euler method steps. |
| | |
| | 1. compute dx_{i}/dt at the current timestep |
| | 2. get \sigma_{up} and \sigma_{down} from ancestral method |
| | 3. compute x_{t-1} = x_{t} + dx_{t}/dt * \sigma_{down} |
| | 4. Add additional noise after the update step x_{t-1} =x_{t-1} + z * \sigma_{up} |
| | """ |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | |
| | sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta) |
| | if callback is not None: |
| | callback({'x': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | |
| | d = to_d(action, sigmas[i], denoised) |
| | |
| | dt = sigma_down - sigmas[i] |
| | |
| | action = action + d * dt |
| | if sigma_down > 0: |
| | action = action + torch.randn_like(action) * sigma_up |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_heun( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | s_churn=0., |
| | s_tmin=0., |
| | s_tmax=float('inf'), |
| | s_noise=1. |
| | ): |
| | """ |
| | Implements Algorithm 2 (Heun steps) from Karras et al. (2022). |
| | For S_churn =0 this is an ODE solver otherwise SDE |
| | Every update consists of these substeps: |
| | 1. Addition of noise given the factor eps |
| | 2. Solving the ODE dx/dt at timestep t using the score model |
| | 3. Take Euler step from t -> t+1 to get x_{i+1} |
| | 4. 2nd order correction step to get x_{i+1}^{(2)} |
| | |
| | In contrast to the Euler variant, this variant computes a 2nd order correction step. |
| | """ |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0. |
| | eps = torch.randn_like(action) * s_noise |
| | sigma_hat = sigmas[i] * (gamma + 1) |
| | |
| | if gamma > 0: |
| | action= action+ eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5 |
| | denoised = model(state, action, goal, sigma_hat * s_in, **extra_args) |
| | d = to_d(action, sigma_hat, denoised) |
| | if callback is not None: |
| | callback({'x': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised}) |
| | dt = sigmas[i + 1] - sigma_hat |
| | |
| | if sigmas[i + 1] == 0: |
| | |
| | action= action+ d * dt |
| | else: |
| | |
| | action_2 = action+ d * dt |
| | denoised_2 = model(state, action_2, goal, sigmas[i + 1] * s_in,**extra_args) |
| | d_2 = to_d( action_2, sigmas[i + 1], denoised_2) |
| | d_prime = (d + d_2) / 2 |
| | action= action+ d_prime * dt |
| | |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpm_2( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | s_churn=0., |
| | s_tmin=0., |
| | s_tmax=float('inf'), |
| | s_noise=1. |
| | ): |
| | """ |
| | A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022). |
| | SDE for S_churn!=0 and ODE otherwise |
| | |
| | 1. |
| | |
| | Last denoising step is an Euler step |
| | """ |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | |
| | gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0. |
| | |
| | eps = torch.randn_like(action) * s_noise |
| | sigma_hat = sigmas[i] * (gamma + 1) |
| | |
| | if gamma > 0: |
| | action = action + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5 |
| | |
| | denoised = model(state, action, goal, sigma_hat * s_in, **extra_args) |
| | d = to_d(action, sigma_hat, denoised) |
| |
|
| | if callback is not None: |
| | callback({'action': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised}) |
| | |
| | |
| | if sigmas[i + 1] == 0: |
| | |
| | dt = sigmas[i + 1] - sigma_hat |
| | action = action + d * dt |
| | else: |
| | |
| | sigma_mid = sigma_hat.log().lerp(sigmas[i + 1].log(), 0.5).exp() |
| | dt_1 = sigma_mid - sigma_hat |
| | dt_2 = sigmas[i + 1] - sigma_hat |
| | action_2 = action + d * dt_1 |
| | denoised_2 = model(state, action_2, goal, sigma_mid * s_in, **extra_args) |
| | d_2 = to_d( action_2, sigma_mid, denoised_2) |
| | action = action + d_2 * dt_2 |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpm_2_ancestral(model, state, action, goal, sigmas, scaler=None, extra_args=None, callback=None, disable=None, eta=1.): |
| | """ |
| | Ancestral sampling with DPM-Solver inspired second-order steps. |
| | |
| | Ancestral sampling is based on the DDPM paper (https://arxiv.org/abs/2006.11239) generation process. |
| | Song et al. (2021) show that ancestral sampling can be used to improve the performance of DDPM for its SDE formulation. |
| | |
| | 1. Compute dx_{i}/dt at the current timestep |
| | |
| | """ |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta) |
| | if callback is not None: |
| | callback({'x': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | d = to_d(action, sigmas[i], denoised) |
| | if sigma_down == 0: |
| | |
| | dt = sigma_down - sigmas[i] |
| | action= action+ d * dt |
| | else: |
| | |
| | sigma_mid = sigmas[i].log().lerp(sigma_down.log(), 0.5).exp() |
| | dt_1 = sigma_mid - sigmas[i] |
| | dt_2 = sigma_down - sigmas[i] |
| | action_2 = action+ d * dt_1 |
| | denoised_2 = model(state, action_2, goal, sigma_mid * s_in, **extra_args) |
| | d_2 = to_d( action_2, sigma_mid, denoised_2) |
| | action= action+ d_2 * dt_2 |
| | action= action+ torch.randn_like(action) * sigma_up |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | def linear_multistep_coeff(order, t, i, j): |
| | ''' |
| | Returns the coefficient of the j-th derivative of the i-th step of a linear multistep method. |
| | ''' |
| | if order - 1 > i: |
| | raise ValueError(f'Order {order} too high for step {i}') |
| | def fn(tau): |
| | prod = 1. |
| | for k in range(order): |
| | if j == k: |
| | continue |
| | prod *= (tau - t[i - k]) / (t[i - j] - t[i - k]) |
| | return prod |
| | return integrate.quad(fn, t[i], t[i + 1], epsrel=1e-4)[0] |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_lms( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | order=4 |
| | ): |
| | ''' |
| | A linear multistep sampler. |
| | |
| | 1. compute x_{t-1} using the current noise level |
| | 2. compute dx/dt at x_{t-1} using the current noise level |
| | ''' |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | sigmas_cpu = sigmas.detach().cpu().numpy() |
| | ds = [] |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | d = to_d(action, sigmas[i], denoised) |
| | ds.append(d) |
| | if len(ds) > order: |
| | ds.pop(0) |
| | if callback is not None: |
| | callback({'x': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | cur_order = min(i + 1, order) |
| | coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)] |
| | action = action + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds))) |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def log_likelihood(model, state, action, goal, sigma_min, sigma_max, extra_args=None, atol=1e-4, rtol=1e-4): |
| | ''' |
| | Computes the log-likelihood of actions |
| | ''' |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | v = torch.randint_like(action, 2) * 2 - 1 |
| | fevals = 0 |
| | def ode_fn(sigma, action): |
| | nonlocal fevals |
| | with torch.enable_grad(): |
| | action= action[0].detach().requires_grad_() |
| | denoised = model(state, action, goal, sigma * s_in, **extra_args) |
| | d = to_d(action, sigma, denoised) |
| | fevals += 1 |
| | grad = torch.autograd.grad((d * v).sum(), action)[0] |
| | d_ll = (v * grad).flatten(1).sum(1) |
| | return d.detach(), d_ll |
| | action_min = action, action.new_zeros([action.shape[0]]) |
| | t = action.new_tensor([sigma_min, sigma_max]) |
| | sol = odeint(ode_fn, action_min, t, atol=atol, rtol=rtol, method='dopri5') |
| | latent, delta_ll = sol[0][-1], sol[1][-1] |
| | ll_prior = torch.distributions.Normal(0, sigma_max).log_prob(latent).flatten(1).sum(1) |
| | return ll_prior + delta_ll, {'fevals': fevals} |
| |
|
| |
|
| | class PIDStepSizeController: |
| | """A PID controller for ODE adaptive step size control.""" |
| | def __init__(self, h, pcoeff, icoeff, dcoeff, order=1, accept_safety=0.81, eps=1e-8): |
| | self.h = h |
| | self.b1 = (pcoeff + icoeff + dcoeff) / order |
| | self.b2 = -(pcoeff + 2 * dcoeff) / order |
| | self.b3 = dcoeff / order |
| | self.accept_safety = accept_safety |
| | self.eps = eps |
| | self.errs = [] |
| |
|
| | def limiter(self, action): |
| | return 1 + math.atan(action- 1) |
| |
|
| | def propose_step(self, error): |
| | inv_error = 1 / (float(error) + self.eps) |
| | if not self.errs: |
| | self.errs = [inv_error, inv_error, inv_error] |
| | self.errs[0] = inv_error |
| | factor = self.errs[0] ** self.b1 * self.errs[1] ** self.b2 * self.errs[2] ** self.b3 |
| | factor = self.limiter(factor) |
| | accept = factor >= self.accept_safety |
| | if accept: |
| | self.errs[2] = self.errs[1] |
| | self.errs[1] = self.errs[0] |
| | self.h *= factor |
| | return accept |
| |
|
| |
|
| | class DPMSolver(nn.Module): |
| | """DPM-Solver. See https://arxiv.org/abs/2206.00927.""" |
| |
|
| | def __init__(self, model, extra_args=None, eps_callback=None, info_callback=None): |
| | super().__init__() |
| | self.model = model |
| | self.extra_args = {} if extra_args is None else extra_args |
| | self.eps_callback = eps_callback |
| | self.info_callback = info_callback |
| |
|
| | def t(self, sigma): |
| | return -sigma.log() |
| |
|
| | def sigma(self, t): |
| | return t.neg().exp() |
| |
|
| | def eps(self, eps_cache, key, state, action, goal, t, *args, **kwargs): |
| | if key in eps_cache: |
| | return eps_cache[key], eps_cache |
| | sigma = self.sigma(t) * action.new_ones([action.shape[0]]) |
| | eps = (action - self.model(state, action, goal, sigma, *args, **self.extra_args, **kwargs)) / self.sigma(t) |
| | if self.eps_callback is not None: |
| | self.eps_callback() |
| | return eps, {key: eps, **eps_cache} |
| |
|
| | def dpm_solver_1_step(self, state, action, goal, t, t_next, eps_cache=None): |
| | eps_cache = {} if eps_cache is None else eps_cache |
| | h = t_next - t |
| | eps, eps_cache = self.eps(eps_cache, 'eps', state, action, goal, t) |
| | action_1 = action- self.sigma(t_next) * h.expm1() * eps |
| | return action_1, eps_cache |
| |
|
| | def dpm_solver_2_step(self, state, action, goal, t, t_next, r1=1 / 2, eps_cache=None): |
| | eps_cache = {} if eps_cache is None else eps_cache |
| | h = t_next - t |
| | eps, eps_cache = self.eps(eps_cache, 'eps', state, action, goal, t) |
| | s1 = t + r1 * h |
| | u1 = action - self.sigma(s1) * (r1 * h).expm1() * eps |
| | eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', state, u1, goal, s1) |
| | action_2 = action - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / (2 * r1) * h.expm1() * (eps_r1 - eps) |
| | return action_2, eps_cache |
| |
|
| | def dpm_solver_3_step(self, state, action, goal, t, t_next, r1=1 / 3, r2=2 / 3, eps_cache=None): |
| | eps_cache = {} if eps_cache is None else eps_cache |
| | h = t_next - t |
| | eps, eps_cache = self.eps(eps_cache, 'eps', state, action, goal, t) |
| | s1 = t + r1 * h |
| | s2 = t + r2 * h |
| | u1 = action - self.sigma(s1) * (r1 * h).expm1() * eps |
| | eps_r1, eps_cache = self.eps(eps_cache, 'eps_r1', state, u1, goal, s1) |
| | u2 = action - self.sigma(s2) * (r2 * h).expm1() * eps - self.sigma(s2) * (r2 / r1) * ((r2 * h).expm1() / (r2 * h) - 1) * (eps_r1 - eps) |
| | eps_r2, eps_cache = self.eps(eps_cache, 'eps_r2', state, u2, goal, s2) |
| | action_3 = action - self.sigma(t_next) * h.expm1() * eps - self.sigma(t_next) / r2 * (h.expm1() / h - 1) * (eps_r2 - eps) |
| | return action_3, eps_cache |
| |
|
| | def dpm_solver_fast(self, state, action, goal, t_start, t_end, nfe, eta=0., s_noise=1., noise_sampler=None): |
| | noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler |
| | if not t_end > t_start and eta: |
| | raise ValueError('eta must be 0 for reverse sampling') |
| |
|
| | m = math.floor(nfe / 3) + 1 |
| | ts = torch.linspace(t_start, t_end, m + 1, device=action.device) |
| |
|
| | if nfe % 3 == 0: |
| | orders = [3] * (m - 2) + [2, 1] |
| | else: |
| | orders = [3] * (m - 1) + [nfe % 3] |
| |
|
| | for i in range(len(orders)): |
| | eps_cache = {} |
| | t, t_next = ts[i], ts[i + 1] |
| | if eta: |
| | sd, su = get_ancestral_step(self.sigma(t), self.sigma(t_next), eta) |
| | t_next_ = torch.minimum(t_end, self.t(sd)) |
| | su = (self.sigma(t_next) ** 2 - self.sigma(t_next_) ** 2) ** 0.5 |
| | else: |
| | t_next_, su = t_next, 0. |
| |
|
| | eps, eps_cache = self.eps(eps_cache, 'eps', state, action, goal, t) |
| | denoised = action- self.sigma(t) * eps |
| | if self.info_callback is not None: |
| | self.info_callback({'x': action, 'i': i, 't': ts[i], 't_up': t, 'denoised': denoised}) |
| |
|
| | if orders[i] == 1: |
| | action, eps_cache = self.dpm_solver_1_step(state, action, goal, t, t_next_, eps_cache=eps_cache) |
| | elif orders[i] == 2: |
| | action, eps_cache = self.dpm_solver_2_step(state, action, goal, t, t_next_, eps_cache=eps_cache) |
| | else: |
| | action, eps_cache = self.dpm_solver_3_step(state, action, goal, t, t_next_, eps_cache=eps_cache) |
| |
|
| | action= action+ su * s_noise * noise_sampler(self.sigma(t), self.sigma(t_next)) |
| |
|
| | return action |
| |
|
| | def dpm_solver_adaptive(self, state, action, goal, t_start, t_end, order=3, rtol=0.05, atol=0.0078, h_init=0.05, pcoeff=0., icoeff=1., dcoeff=0., accept_safety=0.81, eta=0., s_noise=1.): |
| | noise_sampler = default_noise_sampler(action) if noise_sampler is None else noise_sampler |
| | if order not in {2, 3}: |
| | raise ValueError('order should be 2 or 3') |
| | forward = t_end > t_start |
| | if not forward and eta: |
| | raise ValueError('eta must be 0 for reverse sampling') |
| | h_init = abs(h_init) * (1 if forward else -1) |
| | atol = torch.tensor(atol) |
| | rtol = torch.tensor(rtol) |
| | s = t_start |
| | action_prev = action |
| | accept = True |
| | pid = PIDStepSizeController(h_init, pcoeff, icoeff, dcoeff, 1.5 if eta else order, accept_safety) |
| | info = {'steps': 0, 'nfe': 0, 'n_accept': 0, 'n_reject': 0} |
| |
|
| | while s < t_end - 1e-5 if forward else s > t_end + 1e-5: |
| | eps_cache = {} |
| | t = torch.minimum(t_end, s + pid.h) if forward else torch.maximum(t_end, s + pid.h) |
| | if eta: |
| | sd, su = get_ancestral_step(self.sigma(s), self.sigma(t), eta) |
| | t_ = torch.minimum(t_end, self.t(sd)) |
| | su = (self.sigma(t) ** 2 - self.sigma(t_) ** 2) ** 0.5 |
| | else: |
| | t_, su = t, 0. |
| |
|
| | eps, eps_cache = self.eps(eps_cache, 'eps', state, action, goal, s) |
| | denoised = action - self.sigma(s) * eps |
| |
|
| | if order == 2: |
| | action_low, eps_cache = self.dpm_solver_1_step(state, action, goal, s, t_, eps_cache=eps_cache) |
| | action_high, eps_cache = self.dpm_solver_2_step(state, action, goal, s, t_, eps_cache=eps_cache) |
| | else: |
| | action_low, eps_cache = self.dpm_solver_2_step(state, action, goal, s, t_, r1=1 / 3, eps_cache=eps_cache) |
| | action_high, eps_cache = self.dpm_solver_3_step(state, action, goal, s, t_, eps_cache=eps_cache) |
| | delta = torch.maximum(atol, rtol * torch.maximum( action_low.abs(), action_prev.abs())) |
| | error = torch.linalg.norm(( action_low - action_high) / delta) / action.numel() ** 0.5 |
| | accept = pid.propose_step(error) |
| | if accept: |
| | action_prev = action_low |
| | action = action_high + su * s_noise * noise_sampler(self.sigma(s), self.sigma(t)) |
| | s = t |
| | info['n_accept'] += 1 |
| | else: |
| | info['n_reject'] += 1 |
| | info['nfe'] += order |
| | info['steps'] += 1 |
| |
|
| | if self.info_callback is not None: |
| | self.info_callback({'x': action, 'i': info['steps'] - 1, 't': s, 't_up': s, 'denoised': denoised, 'error': error, 'h': pid.h, **info}) |
| |
|
| | return action, info |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpm_fast( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigma_min, |
| | sigma_max, |
| | n, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | eta=0., |
| | s_noise=1., |
| | noise_sampler=None |
| | ): |
| | """DPM-Solver-Fast (fixed step size). See https://arxiv.org/abs/2206.00927.""" |
| | if sigma_min <= 0 or sigma_max <= 0: |
| | raise ValueError('sigma_min and sigma_maactionmust not be 0') |
| | with tqdm(total=n, disable=disable) as pbar: |
| | dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update) |
| | if callback is not None: |
| | dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info}) |
| | return dpm_solver.dpm_solver_fast(state, action, goal, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), n, eta, s_noise, noise_sampler) |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpmpp_2m( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None |
| | ): |
| | """DPM-Solver++(2M).""" |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | sigma_fn = lambda t: t.neg().exp() |
| | t_fn = lambda sigma: sigma.log().neg() |
| | old_denoised = None |
| |
|
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | if callback is not None: |
| | callback({'action': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) |
| | h = t_next - t |
| | if old_denoised is None or sigmas[i + 1] == 0: |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised |
| | else: |
| | h_last = t - t_fn(sigmas[i - 1]) |
| | r = h_last / h |
| | denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised_d |
| | old_denoised = denoised |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpmpp_sde( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | eta=1., |
| | s_noise=1., |
| | scaler=None, |
| | noise_sampler=None, |
| | r=1 / 2 |
| | ): |
| | """DPM-Solver++ (stochastic).""" |
| | x = action |
| | sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max() |
| | noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) if noise_sampler is None else noise_sampler |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = x.new_ones([x.shape[0]]) |
| | sigma_fn = lambda t: t.neg().exp() |
| | t_fn = lambda sigma: sigma.log().neg() |
| |
|
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | denoised = model(state, x, goal, sigmas[i] * s_in, **extra_args) |
| | if callback is not None: |
| | callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | if sigmas[i + 1] == 0: |
| | |
| | d = to_d(x, sigmas[i], denoised) |
| | dt = sigmas[i + 1] - sigmas[i] |
| | x = x + d * dt |
| | else: |
| | |
| | t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) |
| | h = t_next - t |
| | s = t + h * r |
| | fac = 1 / (2 * r) |
| |
|
| | |
| | sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(s), eta) |
| | s_ = t_fn(sd) |
| | x_2 = (sigma_fn(s_) / sigma_fn(t)) * x - (t - s_).expm1() * denoised |
| | x_2 = x_2 + noise_sampler(sigma_fn(t), sigma_fn(s)) * s_noise * su |
| | denoised_2 = model(state, x_2, goal, sigma_fn(s) * s_in, **extra_args) |
| |
|
| | |
| | sd, su = get_ancestral_step(sigma_fn(t), sigma_fn(t_next), eta) |
| | t_next_ = t_fn(sd) |
| | denoised_d = (1 - fac) * denoised + fac * denoised_2 |
| | x = (sigma_fn(t_next_) / sigma_fn(t)) * x - (t - t_next_).expm1() * denoised_d |
| | x = x + noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * su |
| | if scaler is not None: |
| | x = scaler.clip_output(x) |
| | return x |
| |
|
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpmpp_2_with_lms( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None |
| | ): |
| | """DPM-Solver++(2M).""" |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | sigma_fn = lambda t: t.neg().exp() |
| | t_fn = lambda sigma: sigma.log().neg() |
| | old_denoised = None |
| |
|
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | if callback is not None: |
| | callback({'action': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) |
| | h = t_next - t |
| | if old_denoised is None or sigmas[i + 1] == 0: |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised |
| | else: |
| | h_last = t - t_fn(sigmas[i - 1]) |
| | r = h_last / h |
| | denoised_d = (1 + 1 / (2 * r)) * denoised - (1 / (2 * r)) * old_denoised |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised_d |
| | old_denoised = denoised |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpm_adaptive( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigma_min, |
| | sigma_max, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | order=3, |
| | rtol=0.05, |
| | atol=0.0078, |
| | h_init=0.05, |
| | pcoeff=0., |
| | icoeff=1., |
| | dcoeff=0., |
| | accept_safety=0.81, |
| | eta=0., |
| | s_noise=1., |
| | return_info=False |
| | ): |
| | """ |
| | DPM-Solver-12 and 23 (adaptive step size). |
| | |
| | See https://arxiv.org/abs/2206.00927. |
| | """ |
| | if sigma_min <= 0 or sigma_max <= 0: |
| | raise ValueError('sigma_min and sigma_max action nmust not be 0') |
| | with tqdm(disable=disable) as pbar: |
| | dpm_solver = DPMSolver(model, extra_args, eps_callback=pbar.update) |
| | if callback is not None: |
| | dpm_solver.info_callback = lambda info: callback({'sigma': dpm_solver.sigma(info['t']), 'sigma_hat': dpm_solver.sigma(info['t_up']), **info}) |
| | action, info = dpm_solver.dpm_solver_adaptive(state, action, goal, dpm_solver.t(torch.tensor(sigma_max)), dpm_solver.t(torch.tensor(sigma_min)), order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise) |
| | if return_info: |
| | return action, info |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpmpp_2s_ancestral( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | eta=1., |
| | s_noise=1., |
| | noise_sampler=None |
| | ): |
| | """ |
| | Ancestral sampling combined with DPM-Solver++(2S) second-order steps.""" |
| | extra_args = {} if extra_args is None else extra_args |
| | noise_sampler = default_noise_sampler(action) if noise_sampler is None else noise_sampler |
| | s_in = action.new_ones([action.shape[0]]) |
| | sigma_fn = lambda t: t.neg().exp() |
| | t_fn = lambda sigma: sigma.log().neg() |
| |
|
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta) |
| | if callback is not None: |
| | callback({'action': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | if sigma_down == 0: |
| | |
| | d = to_d(action, sigmas[i], denoised) |
| | dt = sigma_down - sigmas[i] |
| | action = action + d * dt |
| | else: |
| | |
| | t, t_next = t_fn(sigmas[i]), t_fn(sigma_down) |
| | r = 1 / 2 |
| | h = t_next - t |
| | s = t + r * h |
| | x_2 = (sigma_fn(s) / sigma_fn(t)) * action - (-h * r).expm1() * denoised |
| | denoised_2 = model(state, x_2, goal, sigma_fn(s) * s_in, **extra_args) |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised_2 |
| | |
| | action = action + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_ddim( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | eta=1., |
| | ): |
| | """ |
| | DPM-Solver 1( or DDIM sampler""" |
| | extra_args = {} if extra_args is None else extra_args |
| | s_in = action.new_ones([action.shape[0]]) |
| | sigma_fn = lambda t: t.neg().exp() |
| | t_fn = lambda sigma: sigma.log().neg() |
| | old_denoised = None |
| |
|
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | if callback is not None: |
| | callback({'action': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) |
| | h = t_next - t |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised |
| | return action |
| |
|
| |
|
| |
|
| | @torch.no_grad() |
| | def sample_dpmpp_2s( |
| | model, |
| | state, |
| | action, |
| | goal, |
| | sigmas, |
| | scaler=None, |
| | extra_args=None, |
| | callback=None, |
| | disable=None, |
| | eta=1., |
| | ): |
| | """ |
| | DPM-Solver++(2S) second-order steps.""" |
| | extra_args = {} if extra_args is None else extra_args |
| | sigma_fn = lambda t: t.neg().exp() |
| | t_fn = lambda sigma: sigma.log().neg() |
| | s_in = action.new_ones([action.shape[0]]) |
| | for i in trange(len(sigmas) - 1, disable=disable): |
| | denoised = model(state, action, goal, sigmas[i] * s_in, **extra_args) |
| | if callback is not None: |
| | callback({'action': action, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) |
| | if sigmas[i + 1] == 0: |
| | |
| | d = to_d(action, sigmas[i], denoised) |
| | dt = sigmas[i + 1] - sigmas[i] |
| | action = action + d * dt |
| | else: |
| | |
| | t, t_next = t_fn(sigmas[i]), t_fn(sigmas[i + 1]) |
| | r = 1 / 2 |
| | h = t_next - t |
| | s = t + r * h |
| | x_2 = (sigma_fn(s) / sigma_fn(t)) * action - (-h * r).expm1() * denoised |
| | denoised_2 = model(state, x_2, goal, sigma_fn(s) * s_in, **extra_args) |
| | action = (sigma_fn(t_next) / sigma_fn(t)) * action - (-h).expm1() * denoised_2 |
| | if scaler is not None: |
| | action = scaler.clip_output(action) |
| | return action |
| |
|
| |
|
| | def make_sample_contour_plot(actions, n_steps, file_store_path): |
| | |
| | store_path = os.path.join(file_store_path, 'action_visualization.png') |
| | rows = n_steps % 2 |
| | for idx, step in range(n_steps): |
| | fig, axs = plt.subplots() |
| | actions |
| | store_path = os.path.join(file_store_path, f'action_visualization_step_{idx}.png') |
| | |
| | |
| | |